On the Vulnerability of the Proportional Fairness Scheduler to Retransmissions Attacks

Udi Ben-Porat ETH Zurich SwitzerlandAnat Bremler-Barr IDC Herzliya IsraelHanoch Levy Tel-Aviv University IsraelBernhard Plattner ETH Zurich Switzerland

INFOCOM 2011

Wireless Scheduling

Players: Base Station (BS) and Clients Time is divided to time slots

In each slot, data is sent to one user only Clients have variable channel conditions The scheduler selects a client for transmission

based on the reported channel condition of the users

2

Scheduler’s objectives: Good overall throughput performance of the system Fairness among users (avoid starvation)

The Common Scheduler – Proportional Fairness (PFS) Fair & Efficient (widely studied and deployed)

PROBLEM: Retransmission policy overlooked ISSUE: Malicious Users can Downgrade Performance Contributions:

Expose vulnerability of the PFS wireless scheduler Examine and Analyze potential solutions Propose a solution that maintains fairness and immune to

attacks

3

* All claims in this work are analytically proved and backed up by simulations.

PROBLEM & Contributions

Proportional Fairness Scheduler (PFS) – User Info

Vi(t)Ri(t)Ai(t)

Throughput Average(until t)

Priority Value

Rate (at time t)

The user with the highest priority is scheduled

Proportional Fairness Scheduler (PFS)

Base Station

DATA

User 1 User 25

V1(t) = 2Priority Value

RateR1(t) = 400 b/s R2(t) = 300 b/s

Rate

V2(t) = 3Priority Value

A1(t) =200A2(t) =100

Throughput Average

Throughput Average Ai(t)

Throughput Average update – “Admitted Average”:

1ircv(t) = 1 if user i received a transmission in time t (o/w 0)

Ri(t) is the “price” the user “pays” per transmission Higher “price” Higher Ai(t) Harder to “win” future time slots

6

Ai(t+1) = (1- ε)Ai(t) + ε1ircv(t)Ri(t)

Frame Losses and Retransmissions

DATA

Sorry, I didn’t get it.Send again please !OK, just

a moment

Hey!What

about us?

Hmm…

1. When to retransmit a lost frame?- Should pending retransmissions get the highest priority?

2. What is the real received data rate?- Due to losses, Ri(t) does not reflect the real rate to the user

8

Frame Losses and Retransmissions

When to retransmit a lost frame? „Fast Ret.“

- Retransmit immediately (ignore other users)

„Slow Ret.“- Some other user has higher priority? - Delay retransmission

Effective Rate - Rei(t)

Effective Rate = The rate the user is expected to receive Example: Ri(t) = 200 Kb/s , Loss Prob. = 0.2 Re

i(t) = 160 Kb/s

Vi(t) = Rei(t) / Ai(t)

Frame Losses and Retransmissions

NEW DATA

200Kb/slot

Regular User

Time Data ack/nak Payment

1

2

101000100101010110111000

= 0200Kb 0 Kb

t = 1

Rate selected: 200 b/sProb. For frame loss = 0.2Re

i(t) = 200(1-0.2)=160 b/sBase Station User i

9

Frame Losses and Retransmissions

RETRANSMISSION

200Kb/slot

t = 2

= 200

200 Kb200Kb

Regular User

Time Data ack/nak Payment

1

2

101000100101010110111000

200Kb 0 Kb

Base Station User i

Rate selected: 200 b/sProb. For frame loss = 0.2Re

i(t) = 200(1-0.2)=160 b/s

10

Our Contributions

1. Retransmissions Scheduling – Vulnerable!2. Propose Immune & Fair solutions

11

Fast Ret. Slow Ret.Averaging Method Immune Fair Immune Fair

Admitted Avg. (original) Transmission Avg. Effective Avg. Initial Effective Avg.

Admitted Average – Malicious Attack

NEW DATA

200Kb/slot= 0

Malicious user

Time Data Ack/Nak Payment

1

2

3

4

5

101000100101010110111000

200Kb 0 Kb

Rate selected: 200 b/sProb. For frame loss = 0.2Re

i(t) = 200(1-0.2)=160 b/s

Base Station MaliciousUser

12

Admitted Average – Malicious Attack

RETRANSMISSION

200Kb/slot= 200

Malicious user

Time Data Ack/Nak Payment

1

2

3

4

5

101000100101010110111000

0 Kb

200Kb 0 Kb

200Kb

200Kb

200Kb

200Kb

0 Kb

0 Kb

200 KbMalicious

UserBase Station

Rate selected: 200 b/sProb. For frame loss = 0.2Re

i(t) = 200(1-0.2)=160 b/s

13

Retransmissions Attack – Simulation Results

Example:- 10% are malicious- # of retransmissions are

limited to Lmax =10

40% time share loss for every regular user

14

X – Percentage of malicious usersY – Time share loss for regular users

Results Fast Ret. Slow Ret.

Averaging Method Immune Fair Immune Fair

Admitted Avg. (original)

15

Sol #1 – Transmission Average

Sol #1 - “Pay” for every transmitted frame

1isnd(t) = 1 if a frame was sent (o/w 0)

16

Ai(t+1) = (1- ε)Ai(t) + ε1isnd(t)Ri(t)

Immunity of “Transmission Average”Malicious user

Time Data Ack/Nak Payment

1

2

3

4

5

101000100101010110111000

200Kb

200Kb

200Kb 200 Kb

200 Kb

200Kb 200 Kb

200Kb

200 Kb

200 Kb

Total “Payment”: 1000Kb

200 Kb200Kb

Regular User

Time Data Ack/Nak Payment

1

2

101000100101010110111000

200Kb 200 Kb

Total “Payment”: 400Kb

The Scheduler is immune to attack

What about Fairness?

17

Transmission Average – Distorted Fairness

18

2 Regular users Results*

UserEffective

RateRe

i(t)

Rate

Ri(t)

Frame Loss Prob. “Payment” per bit

Time Share

Download Speed

A 100 b/s 200 b/s 0.5 1/(1-0.5) = 2 38% 38 b/s

B 90 b/s 100 b/s 0.1 1/(1-0.1) = 1.1 62% 55.8 b/s

User A has a better channel condition than User B, but still… …gets smaller time share …receives less throughput

This stands against any notion of fairness!

* Long run results. Derived from an analytical result proved in the paper.

Results Fast Ret. Slow Ret.

Averaging Method Immune Fair Immune Fair

Admitted Avg. (original) Transmission Avg.

19

Sol #2 – Effective Average

“Pay for what I expect you to receive”

Immune: Malicious has to pay for excessive ret. Fair: Ai(t+1) = The throughput user i actually received

20

Ai(t+1) = (1- ε)Ai(t) + ε1isnd(t)Re

i(t)

Effective Average (Sol. #2)

Regular user

TimeRe

i(t) Effective Rate Data ACK/NAK Payment

1

3

101000100101010110111000

300 b

300 b

150 b/s

150 b/s

150 b/s

150 b/s

Regular user: 300 b/s, 2 transmissions

Rate selected: 300 b/sFrame loss Probability = 1/2

Rei(t) = 150 b/s

21

Effective Average (Sol. #2) for Fast retransmissions

Malicious user

TimeRe

i(t) Effective Rate Data ACK/NAK Payment

1

2

3

Malicious user: 170 b/s, 3 transmissions

101000100101010110111000

300 b

10 b/s

150 b/s

10 b/s

Rate selected: 300 b/sFrame loss Probability = 1/2

Rei(t) = 150 b/s

10 b/s

150 b/s

10 b/s

Regular user: 300 b/s, 2 transmissions

22

Results Fast Ret. Slow Ret.

Averaging Method Immune Fair Immune Fair

Admitted Avg. (original) Transmission Avg. Effective Avg.

23

Sol #3 – Initial Effective Rate (for Fast Ret.)

Initial Effective Rate (sol #3) Every retransmission costs as the first transmission

Fi (t)= The last time slot where user i received an initial trans.

Choosing fast retransmission is preferred when: Small changes in channel conditions between slots

- Time slots are very short- Channel condition is stable

The preferred user in time t, is probably also the one in t+1 24

Ai(t+1) = (1- ε)Ai(t) + ε1isnd(t) Re

i(Fi (t))

Results Fast Ret. Slow Ret.

Averaging Method Immune Fair Immune Fair

Admitted Avg. (original) Transmission Avg. Effective Avg. Initial Effective Avg.

25

Conclusions

The Proportional Fairness Scheduler is vulnerable to retransmissions attacks Both for Fast and Slow retransmission methods

We proposed modifications to PFS Proved to be proportional fair and immune to ret. attacks Both for Fast and Slow retransmission methods

26

Questions?

27